Jersey number detection using synthetic data in a low-data regime

Area of Science:

• Computer Vision
• Sports Analytics
• Machine Learning

Background:

• Player identification in sports video analysis is crucial but challenging.
• Automatic jersey number detection faces issues like varying camera angles, low resolution, and player movement.

Purpose of the Study:

• To develop a novel approach for jersey number identification in sports videos.
• To address data scarcity and imbalance in datasets using synthetic data generation.

Main Methods:

• Generated synthetic datasets to mitigate data imbalance.
• Employed a multi-step strategy involving person detection and human pose estimation.
• Utilized Convolutional Neural Networks (CNNs) with multi-class and multi-label objectives for number identification.

Main Results:

• Synthetic data generation improved CNN model accuracy by 9% overall.
• Accuracy for low-frequency jersey numbers increased by 18%.

Conclusions:

• The proposed synthetic data generation and pose-guided localization method effectively enhances jersey number identification.
• This approach overcomes limitations of traditional methods and improves performance on imbalanced datasets.